Task-Adaptive Wasserstein Barycentric Fusion for Multimodal Medical Imaging

WBVAE-TA achieves mean Dice scores of 62.31%, 78.83%, and 87.30% for enhancing tumor, tumor core, and whole tumor on BraTS 2018 across all 15 MRI modality combinations.

• Why it matters: Multimodal fusion must reflect task-dependent modality contributions while preserving shared and modality-specific information, especially under complete and incomplete MRI inputs.
• What they did: The Task-Adaptive Wasserstein Barycenter Variational Autoencoder learns modality weights jointly with the application objective while aggregating diagonal Gaussian encodings via a Wasserstein barycenter.
• The result: In normative modeling on UK Biobank and ADNI, it attains the highest estimated data log-likelihood and the highest ADNI significance ratio, precision, and balanced accuracy among evaluated methods.